Pre-written path. Predictable, cheap, testable. Fails when the input varies.
Model chooses the path. Handles variety and judgment. Costs more to run and to trust.
A desk: workflow skeleton, agents at the judgment points, a person at the gate.
Deterministic wherever possible. Agents only where judgment is required.
The difference in one line each
Workflow: the code decides. Every step and branch was written by a person in advance. Given the same input it does the same thing every time. This is what Zapier, Make, n8n, and most RPA do, and what most of any good AI system should be.
Agent: the model decides. It is given a goal and a set of tools, reads what comes back from each tool call, and chooses the next step itself. This is what handles a vendor's paperwork arriving as four emails and two PDFs in no particular order.
| Workflow automation | AI agent | |
|---|---|---|
| Who controls the path | The code, written in advance | The model, at run time |
| Input | Structured, predictable | Messy, variable, human |
| Cost to run | Near zero | Model usage per run |
| Testing | Unit tests | Evaluation against a golden set |
| Failure mode | Breaks loudly on unexpected input | Produces a plausible wrong answer quietly |
| Trust | Earned once | Earned continuously, needs a gate |
When each is right
Use a workflow when the input is structured and the steps are known: move the file, create the record, send the reminder at day 21, post the total. Do not put a model in the loop for these. It adds cost and a new way to be wrong.
Use an agent when the input varies and a judgment is required: does this certificate cover the scope of work? Which of these three crews fits this order? What is this email actually asking for? A person could answer in seconds; a workflow cannot answer at all.
The combination is what ships
A desk is a workflow skeleton with agents at the judgment points and a person at the gate. Collect (workflow) → extract (agent reads the messy documents) → match (workflow, against your rules) → draft (agent writes it in your voice) → a person approves → out the door (workflow). The deterministic majority is written once and verified; the agents are evaluated against real examples; the gate catches the plausible wrong answer before it leaves.
Vendors that make everything an agent throw away the reliability argument. Vendors that make everything a workflow cannot handle the real input. The split is the design.
Three questions that settle it
- Could a clerk with a checklist do this step? Workflow.
- Would the clerk need to read and decide? Agent, with a typed answer and a source citation.
- Does the result go out under the company's name? A person approves it, whatever produced it.